Context-Specific Independence in Directed Relational Probabilistic Models and its Influence on the Efficiency of Gibbs Sampling